Abstract:
This paper proposed a new emotion recognition system using intrasegmental features, extracted from long monophthongs in continuous speech. 36 vocal tract features and 11 glottal source features were initially extracted and an optimal subset was selected using Maximum Relevance Minimal Redundancy Backward Wrapping (MRMRBW). A newly constructed JL corpus was used to evaluate the system performance. Five different classifiers were considered. By using the optimal classifier, we achieved recognition accuracy of 70.5% regardless of vowel types for five different emotions.